Auto mode can tell Claude to finish work the task exposed

What

Auto mode lets Claude work through a task with fewer permission prompts. A new finishExposedWork option adds instructions to the system prompt, the standing instructions Claude receives. They tell Claude to finish the job the user wants, not just the literal request: run the project's checks, fix failures and bugs the task exposed, and not end a turn by offering more work. Claude should ask the user only for values they must supply and before destructive steps.

• The auto mode reminder now carries finishExposedWork. In the bash-first, steer-only case it comes from the config value finishExposedWork. Otherwise it is true only when the flag tengu_vivid_crown is on (its fallback is off), the session is not in bypass-permissions mode, is not a subagent (a helper copy of Claude), and proactivity is the default. With it on, the reminder can be added where none was added before. Nothing has been read about tengu_vivid_crown, so whether it is on for anyone is unknown.
• When auto mode ends, the exit reminder can now say: 'The auto-mode guidance to fix what the task exposed and not end the turn on an offer no longer applies.' The auto mode exit event also includes finishExposedWork, next to bashFirst and steerOnly.
• When auto mode denies a built-in tool, the record of that denial now includes the command text for certain tools, such as Bash.

Why

When it applies, this could noticeably change how far Claude goes on its own at the end of a turn, fixing what it uncovered instead of stopping to offer. Denials are also easier to understand because the refused command is kept.